Entropy-Based Modeling For Detecting Behavioral Anomalies in Users of a Diabetes Lifestyle Management Support System
Identifying non-adherence indicators in a chatbot-based diabetes support system
Bachelor Thesis
(2025)
Author(s)
Ciuntu (TU Delft - Electrical Engineering, Mathematics and Computer Science)
Contributor(s)
Catholijn Jonker – Mentor (TU Delft - Interactive Intelligence)
J.D. Top – Mentor (TU Delft - Interactive Intelligence)
A. Anand – Graduation committee member (TU Delft - Web Information Systems)
Faculty
Electrical Engineering, Mathematics and Computer Science
To reference this document use:
https://resolver.tudelft.nl/uuid:3cbd29f6-a8ac-4ea9-9b86-1671c4b3af63
More Info
expand_more
expand_more
Publication Year
2025
Language
English
Graduation Date
25-06-2025
Awarding Institution
Delft University of Technology
Project
['CSE3000 Research Project']
Programme
['Computer Science and Engineering']
Faculty
Electrical Engineering, Mathematics and Computer Science
Reuse Rights
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Individuals with diabetes face rigorous demands when it comes to managing their health, yet patients sometimes struggle to stay adherent to treatment. CHIP is an AI-based conversational platform that allows patients to report lifestyle factors and receive personalized support for making healthy lifestyle changes. However, detecting patient non-adherence remains a significant challenge in this system, as this can hinder treatment and complicate decision-making for healthcare providers.
This study presents an anomaly detection system designed to identify behavioral changes in diabetes patients through their chatbot interactions. Such shifts have previously been shown to correlate with non-adherence. The approach extracts temporal, frequency, and content features from patient-chatbot conversations and quantifies behavioral variability using entropy to detect deviations from individual baseline patterns.
The approach was evaluated using synthetic patient-chatbot conversations generated by a locally-hosted large language model, with behavioral shifts manually introduced in the simulated users. The system detected these irregularities with an accuracy of approximately 76% and a recall of around 35%. However, the false positive rate remained high, at around 15%, primarily due to over-flagging in users with naturally high variability. Future improvements could involve machine learning-based personalization to better distinguish between true anomalies and normal variability. With refined detection thresholds, integration into CHIP may enable timely support for patients at risk of non-adherence.